Anomaly Detection in Cyber Security for IoT using Federated Learning

Bargava Subramanian (~bargava)


Description:

The number of IoT devices is expected to reach 50 Billion by 2023. Cyber Security on IoT introduces some significant challenges:

  • Data collected are very sensitive. Most of the times they capture personal data and/or business critical data. Privacy is very important
  • Each device generates data at scale. But the device has constraints on memory, computation and battery life. Connecting each device to the cloud reduces the life of the device.

This calls for a new paradigm in Machine Learning that exhibits the following features:

  • Data should not be moved out of the device and/or network
  • The Machine Learning models should be privacy-preserving

Anomaly detection is a class of unsupervised machine learning models that identifies anomalies in network data. In cybersecurity, anomalies are flagged as a potential threat. In this talk, the speaker discusses how to build anomaly detection models for IoT that satisfies the above two features.

Federated learning is a family of Machine Learning algorithms that has the core idea: a connected network exists in which there is a central server node. Each of the nodes creates data - that has to be used for training as well as for prediction. Each of the nodes trains a local model and only that model is shared with the server, not the data.

But a man-in-the-middle attack can siphon off the data. How to address this? This is done using encryption. The data from the edge node can only be decrypted by the central node.

How will this be privacy preserving if all this does is decentralized learning and encryption? The third and final step to achieve truly privacy-preserving machine learning is to use differential privacy. A common approach to achieve differential privacy is to add random noise to the data before training. The speaker will discuss how to do local and global differential privacy.

In summary, Federated learning enables Edge devices to collaboratively learn a machine learning model but keeping all of the data on the device itself. Federated Learning gives the following advantages:

  • Low latency
  • Privacy-Preserving
  • Energy Efficient

In this talk, the speaker talks how to build anomaly detection models using federated learning on tensorflow. The speaker shows hows to build custom algorithms and loss functions. The inference is done using uTensor - a light weight AI inference library based on mbed and TensorFlow. To deploy models on the devices, the model size has to be small. The speakers discuss briefly on how to achieve that (network compression, quantization etc).

Outline of the talk:

  • Problem Overview: Anomaly detection in Cyber Security for IoT
  • Introduction to Federated Learning
    • Decentralized Training
    • Encryption
    • Differential Privacy
  • Federated Learning using Tensorflow
  • Custom algorithm and loss function
  • Inference using uTensor
  • Deep Learning Model compression for IoT devices
  • Demo

References

https://blog.fastforwardlabs.com/2018/11/14/federated-learning.html

https://federated.withgoogle.com/

https://florian.github.io/federated-learning/

Prerequisites:

Basic understanding of machine learning

Speaker Info:

Bargava Subramanian is a Deep Learning engineer and co-founder of an AI-based Cyber Security startup for IoT, Binaize Labs, in Bangalore, India. He has 15 years’ experience delivering business analytics and machine learning solutions to B2B companies. He mentors organizations in their data science journey. He holds a master’s degree from the University of Maryland at College Park. He is an ardent NBA fan. He is reachable on twitter @bargava

Speaker Links:

https://speakerdeck.com/bargava/

https://medium.com/@bargava

Id: 1432
Section: Data Science, Machine Learning and AI
Type: Talks
Target Audience: Beginner
Last Updated: